Removal of evidential motion-contaminated and poorly fitted image data improves IVIM diffusion MRI parameter scan

Abstract

Background

It has been reported that intravoxel incoherent motion (IVIM) diffusion magnetic resonance imaging (MRI) scan–rescan reproducibility is unsatisfactory.

Purpose

To study IVIM MRI parameter reproducibility for liver parenchyma after the removal of motion-contaminated and/or poorly fitted image data.

Material and Methods

Eighteen healthy volunteers had liver scans twice in the same session to assess scan–rescan repeatability, and again in another session after an average interval of 13 days to assess reproducibility. Diffusion-weighted images were acquired with a 3-T scanner using respiratory-triggered echo-planar sequence and 16 b-values (0–800 s/mm²). Measurement was performed on the right liver with segment-unconstrained least square fitting. Image series with evidential anatomical mismatch, apparent artifacts, and poorly fitted signal intensity vs. b-value curve were excluded. A minimum of three slices was deemed necessary for IVIM parameter estimation.

Results

With a total 54 examinations, six did not satisfy inclusion criteria, leading to a success rate of 89%, and 14 volunteers were finally included for the repeatability/reproducibility study. A total of 3–10 slices per examination (mean = 5.3 slices, median = 5 slices) were utilized for analysis. Using threshold b-value = 80 s/mm², the coefficient of variation and within-subject coefficient of variation for repeatability were 2.86% and 3.36% for D_slow, 3.81% and 4.24% for perfusion fraction (PF), 18.16% and 24.88% for D_fast; and those for reproducibility were 2.48% and 3.24% for D_slow, 4.91% and 5.38% for PF, and 21.18% and 30.89% for D_fast.

Conclusion

Removal of motion-contaminated and/or poorly fitted image data improves IVIM parameter reproducibility.

Keywords

Diffusion-weighted imaging intravoxel incoherent motion liver reproducibility

Introduction

Since the initial study by Yamada et al. (1), there have been greater interests to explore intravoxel incoherent motion (IVIM) diffusion magnetic resonance imaging (MRI) technique to evaluate diffused liver diseases such as liver fibrosis and non-alcoholic fatty liver disease, to characterize liver tumor, and to evaluate treatment response (2,3). A prerequisite to translating IVIM imaging into clinical applications is accurate measurement of IVIM parameters and acceptable reproducibility. Nevertheless, accurate liver IVIM quantification is challenging, partially due to the limited sampling and low signal-to-noise ratio (SNR) for fast diffusion data acquisition. A major obstacle for clinical application of IVIM technique for abdominal organs is its unsatisfactory scan–rescan reproducibility (2,3). For example, in a short-term reproducibility study, Andreou et al. (4) reported the 95% confidence intervals (CI) of percentage difference between paired measurements of liver parenchyma was (−24.3∼25.1) for PF (f), (−5.12∼8.09) for D_slow (D), and (−31.2∼59.1) for D_fast (D*); and the absolute limit was (0.140∼0.232) for f, (0.951∼1.08) for D, and (35.7∼82.5) for D*.

IVIM diffusion imaging typically involves long data acquisition time with images acquired at a series of b-values. It is usually acquired with respiratory gating; however, even respiratory gating is associated with substantial residual respiration induced motion (2). Respiratory motion can cause inter-b-value motion and intra-b-value motion. Inter-b-value motion causes mismatch of anatomical structures on images of difference b-values, and intra-b-value motion when motion occurs during the data acquisition for the slice causes image artifacts. Diffusion MRI is also influenced by artifacts related to magnet/sequence imperfections, such as B₀ inhomogeneity resulting from susceptibility variations and geometric distortions from residual motion probing gradients-induced eddy currents (5). In this study, we introduce a manual “image data cleaning” process, with the aim to mitigate artifacts associated with respiration motion as well as with magnet/sequence imperfections. We hypothesize that if IVIM’s scan–rescan reproducibility will be satisfactory after “image data cleaning,” then with further technical improvement, such as acquisition of more b-values for curve fitting, advanced methods for motion correction, statistical removal of ill fitted pixels, or accelerated data acquisition with single-breath hold, liver IVIM will eventually have clinical diagnostic applicability.

Material and Methods

This study was conducted with the approval of the institutional ethics committee and informed consent was obtained. Eighteen healthy volunteers underwent IVIM diffusion imaging with a 3-T magnet and a 32-channel dStream Torso coil (Ingenia, Philips Healthcare, Best, The Netherlands). The IVIM diffusion imaging was based on a single-shot spin-echo-type echo-planar imaging sequence, with 16 b-values of 0, 3, 10, 25, 30, 40, 45, 50, 80, 200, 300, 400, 500, 600, 700, and 800 s/mm², NSA of 2 for b = 700 s/mm² and b = 800 s/mm², and NSA = 1 for other b-values. Spectral pre-saturation with inversion recovery technique was used for fat suppression. Respiratory triggering was performed using an air-filled pressure sensor fixed on the upper abdomen, resulting in an average TR of 2149 ms. Other parameters included TE = 55 ms, slice thickness = 6 mm, matrix = 100 × 116, field of view (FOV) = 360 × 300 mm, EPI factor = 29, a sensitivity-encoding (SENSE) factor = 4, number of slices = 26. The scan participants were trained so that they maintained gentle regular breathing during the image acquisition. The average IVIM scan duration was 6 min. All volunteers were scanned twice during the same session to assess scan–rescan repeatability (scans 1.1 and scans 1.2) and additionally once again in another session (scan 2) with an interval of 5–21 days (mean = 13 days) to assess scan–rescan reproducibility.

The IVIM signal attenuation was modeled according to Eq. 1:

SI (b) = {SI}_{0} [(1 - PF) \cdot exp (- b \cdot D_{slow}) + PF \cdot exp (- b \cdot D_{fast})]

(1)

where SI(b) and SI₀ denote the signal intensity acquired with the b-factor value of b and b = 0 s/mm², respectively (6). Perfusion fraction (PF) is the fraction of the pseudo-diffusion linked to microcirculation, D_slow is the diffusion coefficient representing the pure molecular diffusion, and D_fast is the pseudo-diffusion coefficient representing the incoherent microcirculation within the voxel (perfusion-related diffusion). With the segment-unconstrained approach, the estimation of D_slow was obtained by a least-squares linear fitting of the logarithmized image intensity at b-values ≥threshold, which was chosen as 50, 80, or 200 s/mm², to a linear equation (7). The fitted curve was then extrapolated to obtain an intercept at b-value = 0 s/mm². The ratio between this intercept and SI₀ gave an estimated PF. Finally, the obtained D_slow and PF were substituted into Eq. 1 and were nonlinear least squares fitted against all b-values to estimate D_fast using the Trust-Region-based algorithm in MATLAB. The constraints boundaries for D_fast included 0 ≤D_fast ≤200 × 10^–3 mm²/s and starting point = 10 × 10^–3 mm²/s (2).

All curve-fitting algorithms were implemented in a custom program developed on MATLAB (Mathworks, Natick, MA, USA). First, a pixel-wise analysis algorithm was performed in order to exclude pixels which presented a poor fitting for D_slow with a coefficient of determination R² value < 0.8 (8). Regions of interest (ROIs) were then placed to cover a large portion of right liver parenchyma while avoiding large vessels on D_slow parametric map after the poorly fitted pixel exclusion process. For ROI analysis, the IVIM parameters were calculated based on the mean signal intensity of the whole ROI, which has been shown to offer better estimation than pixel-wise fitting when the SNR of the diffusion-weighted images is low (9,10). As in other reports (11 –19), only the right lobe of the liver was measured in the current study (Fig. 1).

Fig. 1.

One IVIM image series of “good quality.” No evidential motion or artifacts are seen in more than two images. ROI was drawn to cover as much as possible of liver parenchyma of right lobe, while avoiding signal vessels which are noticeable in the central parts of parenchyma and also in the area close to the gallbladder in this slice. The signal intensity and b-value curve have a R²>0.96 for ROI-based fitting. The signal decay is consistent with b-value changes and no obvious outlier value is noted.

A manual procedure was taken to “clean the image data” for each examination. First, slices which covered only the lowest part of segment V-VI (usually slices below the gallbladder) or the hepatic dome, near the digestive tract, or the diaphragmatic surfaces were discarded. Then, each scan’s image series were graded by a radiologist as “good quality,” “fair quality,” or “insufficient quality.” Motion-induced imaging data degrading was visually assessed between consecutives images at different b-values for each slice, noting the location of the following anatomic structures: kidneys, gall bladder, spleen, hepatic edges, main hepatic vessels (main portal vein, portal veins until second order, main hepatic veins). If no motion or artifact was noted, the slice series was graded “good quality” (Fig. 1). Image series of “insufficient quality” were mainly due to motion leading to liver displacement between images of different b-values (inter-b-value motion), and sometimes apparent artifacts in the hepatic parenchyma which could be due to intra-b-value motion (Fig. 2). Slices presented only slight displacement or inconspicuous artifact were graded “fair quality.” Image series of “good quality” and “fair quality” were included for the second step data cleaning.

Fig. 2.

An example of excluded image series due to inter-image motion. Left kidney size varies among different images and the posterior branch of the right portal vein can be seen in some images but not in the other images.

Image series which generated a poor IVIM diffusion fitted curve were then excluded. First, slices which presented parameters results with a coefficient of determination R² value < 0.95 for ROI-wise fitting were excluded (20). Then, the plots of signal intensity vs. b-values were individually evaluated. Slices which demonstrated evidential multiple outliers with MRI signal vs. b-value relation and could not be properly fitted were discarded, as well as those resulted in unreasonable high D_fast value approaching the upper boundary of 200 × 10^–3 mm²/s (Fig. 3). The initial part of the curve associated with low b-values tended to be difficult to fit accurately. Very fast signal decay sometimes occurred at b-value ≤ 10 s/mm² (21,22). This strong signal attenuation, which occurs at b-values < 10 s/mm² and may reflect the velocity of flowing protons present in relatively large vessels instead of the micro-vessels of IVIM theory (22), can lead to overestimation of D_fast. In addition, for an IVIM image series to be valid, we required that at least three slices be included for final analysis after data-cleaning. The mean of all included slice measurements was then regarded as the value of the examination.

Fig. 3.

Examples of well-fitted (a), acceptable (b), and unacceptable (c–f) signal intensity and b-value curves. (c–f) MRI signal increases and decreases erratically between b-value = 0 s/mm² and = 80 s/mm², and at least three consecutive data points are outliers under or above the fitted curve between b = 0 s/mm² and b = 80 s/mm². (c, e) Sharp signal drops at b-value = 0 and = 3 s/mm² which therefore derive unreasonably high D_fast values.

Statistical analysis was performed using MedCalc Statistical Software (version 17.6, MedCalc Software bvba, Ostend, Belgium). Intra-scan repeatability between scan 1.1 and scan 1.2, and inter-scan reproducibility between scan 1.1 and scan 2 of PF, D_slow, and D_fast were assessed by coefficient of variation (CoV), the within-subject coefficient of variation (_wCoV), and Bland–Altman mean difference and 95% limits of agreements (BA-LA). _wCoV is defined by Eqs. 2 and 3:

Within - subjectSD = \sqrt{\frac{\sum (x_{1} - x_{2})^{2}}{2 n}}

(2)

Within - subjectCoV (%) = 100 * \frac{Within - subjectSD}{Mean}

(3)

With n being the number of individuals (=14 in this study) and x₁ and x₂ the duplicate parameter measurements for each individual.

Results

With the total 54 examinations, data-cleaning removed six examinations without satisfying inclusion criteria; leading to a success rate of 89% (Fig. 4). Fourteen volunteers were finally included for measurement reproducibility analysis (five men, nine women; mean age = 25.7 years; age range = 24–27 years). Of the scanned slices, 68.11% were included for the final analysis, with a median of five slices (mean = 5.3 slices, range = 3–10 slices) for each scan. The image series graded of “good quality” presented a higher rate of acceptable fitted curve (81.2%) than slices graded “fair quality” (62.1%).

Fig. 4.

(a) Image-data cleaning process flow diagram. (b) Number of volunteers included for analysis at each step.

The scan–rescan repeatability and reproducibility for PF, D_slow, and D_fast using threshold b-value = 80 s/mm² are summarized in Tables 1 –3. Bland–Altman plots for each parameter are shown in Fig. 5. The 95% BA-LA, mean of CoV and _wCoV for repeatability and reproducibility are shown in Tables 4 and 5. Typically, for threshold b-value = 80 s/mm², D_slow had a CoV and _wCoV of 2.86% and 3.36% between scan 1.1 and scan 1.2, and 2.48% and 3.24% between scan 1.1 and scan 2; PF had a CoV and _wCoV of 3.81% and 4.24% between scan 1.1 and scan 1.2, and 4.91% and 5.38% between scan 1.1 and scan 2; D_fast had a CoV and _wCoV of 18.16% and 24.88% between scan 1.1 and scan 1.2, and 21.18% and 30.89% between scan 1.1 and scan 2.

Fig. 5.

Bland–Altman plots of scan–rescan repeatability (scan 1.1 vs. scan 1.2) and reproducibility (scan 1.1 vs. scan 2).

Table 1.

PF value (%) in healthy volunteers with b-value = 80 s/mm² threshold and the number of included slices per examination.

	PF value (included slices [n], CoV_slices [%])
Case no.	Scan 1.1	Scan 1.2	Scan 2
1	22.23 (5, 13.76)	23.14 (3, 12.98)	22.70 (4, 19.13)
2	25.40 (3, 28.63)	26.93 (4, 25.65)	27.03 (3, 9.08)
3	31.72 (3, 15.98)	33.86 (3, 17.64)	33.63 (4, 2.80)
4	25.20 (6, 14.85)	26.24 (3, 9.48)	23.00 (3, 7.82)
5	22.88 (3, 15.45)	24.69 (5, 6.84)	22.11 (7, 8.54)
6	27.53 (4, 12.44)	25.91 (3, 10.79)	27.91 (3, 15.76)
7	22.25 (4, 16.57)	22.65 (5, 16.08)	21.09 (4, 21.31)
8	20.25 (8, 13.07)	20.46 (9, 13.36)	23.23 (7, 17.88)
9	20.47 (6, 9.36)	21.97 (6, 15.99)	22.38 (5, 7.89)
10	26.38 (6, 13.86)	27.50 (5, 8.85)	24.19 (7, 12.21)
11	18.44 (8, 11.91)	16.87 (7, 32.05)	20.85 (5, 10.83)
12	20.73 (9, 11.61)	18.88 (10, 19.29)	18.86 (9, 11.20)
13	21.93 (6, 12.93)	20.18 (6, 11.37)	20.05 (7, 18.53)
14	19.09 (6, 8.12)	18.93 (4, 14.96)	19.44 (4, 10.15)

CoV_slices: Coefficient of variation of PF across the slices included after full data-cleaning.

Table 2.

D_slow values (×10^–3 mm²/s) in healthy volunteers with b-value = 80 s/mm² threshold.

Case no.	Scan 1.1	Scan 1.2	Scan 2
1	0.96 (7.72)	0.92 (3.36)	1.00 (2.93)
2	1.09 (7.65)	1.07 (4.03)	1.10 (7.66)
3	0.98 (3.41)	0.91 (2.30)	0.95 (3.38)
4	0.98 (8.22)	0.97 (6.27)	0.97 (3.17)
5	1.05 (0.59)	1.01 (5.21)	0.96 (4.68)
6	1.03 (5.12)	1.06 (10.49)	1.03 (1.64)
7	1.03 (6.72)	1.05 (14.31)	0.97 (4.44)
8	0.98 (3.79)	1.01 (7.48)	1.00 (3.05)
9	1.00 (3.93)	1.01 (8.95)	0.97 (2.48)
10	1.01 (7.02)	0.96 (10.27)	1.03 (11.07)
11	0.93 (3.71)	0.97 (3.60)	1.02 (7.59)
12	1.09 (6.28)	1.14 (6.58)	1.02 (3.53)
13	1.00 (1.40)	1.10 (7.34)	0.99 (2.99)
14	0.92 (4.72)	0.95 (4.98)	0.93 (7.11)

CoV_slices (%): coefficient of variation of D_slow across the slices included after full data-cleaning.

Table 3.

D_fast value (included slices [n], CoV_slices [%]).

Case no.	Scan 1.1	Scan 1.1	Scan 2
1	40.02 (19.61)	47.17 (17.58)	38.70 (20.91)
2	50.26 (26.07)	40.92 (8.17)	42.49 (22.08)
3	54.79 (24.04)	40.95 (25.25)	55.28 (12.61)
4	81.80 (46.55)	57.29 (67.71)	82.07 (119.20)
5	60.82 (44.72)	64.12 (57.65)	55.76 (21.60)
6	48.91 (28.25)	93.83 (74.85)	100.25 (90.14)
7	32.77 (32.79)	43.36 (37.18)	42.97 (11.26)
8	51.90 (32.92)	63.94 (38.19)	55.95 (31.65)
9	57.68 (24.53)	49.25 (28.48)	71.08 (27.25)
10	97.62 (46.29)	62.33 (34.16)	46.72 (25.86)
11	60.14 (64.20)	66.65 (47.33)	31.76 (29.84)
12	72.91 (42.31)	40.97 (24.16)	46.77 (34.68)
13	47.37 (18.40)	53.80 (31.23)	90.40 (54.72)
14	57.04 (41.15)	58.62 (24.75)	74.67 (56.94)

CoV_slices (in %): coefficient of variation of D_fast across the slices included after full data-cleaning.

Table 4.

Scan–rescan repeatability of IVIM parameters.

		threshold b=50 (s/mm²)	threshold b=80 (s/mm²)	threshold b=200 (s/mm²)
PF	Mean CoV (%, range)	2.40 (0.13 ∼ 7.20)	3.81 (0.62 ∼ 6.61)	4.89 (0.00 ∼ 11.49)
	Within-subject CoV (%)	3.20	4.24	6.32
	Mean Diff % (BA – LA, %)	-0.9 (-10.1 ∼ 8.3)	-0.6 (-12.8 ∼ 11.6)	-3.0 (-19.1 ∼ 13.0)
D_slow	Mean CoV (%, range)	2.07 (0.30 ∼ 5.10)	2.86 (0.43 ∼ 7.18)	3.65 (0.51 ∼ 10.34)
	Within-subject CoV (%)	2.47	3.36	4.43
	Mean Diff % (BA – LA, %)	-0.7 (-7.6 ∼ 6.3)	-0.5 (-10.1 ∼ 9.1)	0.9 (-12.1 ∼ 13.9)
D_fast	Mean CoV (%, range)	23.71 (5.01 ∼ 49.02)	18.16 (1.93 ∼ 44.50)	17.07 (0.47 ∼ 50.39)
	Within-subject CoV (%)	28.28	24.88	26.50
	Mean Diff % (BA – LA, %)	0.9 (-77.2 ∼ 79.0)	3.0 (-60.1 ∼ 66.0)	7.6 (-54.6 ∼ 69.9)

Diff, difference; BA-LA, Bland–Altman 95% limits of agreement.

Table 5.

Scan–rescan reproducibility of IVIM parameters.

		threshold b=50 (s/mm²)	threshold b=80 (s/mm²)	threshold b=200 (s/mm²)
PF	Mean CoV (%, range)	5.68 (0.75 ∼ 10.94)	4.91 (0.97 ∼ 9.71)	5.98 (0.60 ∼ 13.87)
	Within-subject CoV (%)	6.59	5.38	6.94
	Mean Diff % (BA – LA, %)	-0.7 (-19.9 ∼ 18.5)	-0.5 (-16.5 ∼ 15.5)	-1.8 (-22.2 ∼ 18.6)
D_slow	Mean CoV (%, range)	2.99 (0.61 ∼ 6.75)	2.48 (0.05 ∼ 6.45)	3.10 (0.58 ∼ 6.00)
	Within-subject CoV (%)	3.72	3.24	3.53
	Mean Diff % (BA – LA, %)	0.8 (-9.7 ∼ 11.3)	0.8 (-8.4 ∼ 10.0)	1.3 (-8.4 ∼ 11.0)
D_fast	Mean CoV (%, range)	21.60 (1.15 ∼ 67.31)	21.18 (0.23 ∼ 49.87)	19.33 (2.31 ∼ 72.36)
	Within-subject CoV (%)	34.33	30.89	35.93
	Mean Diff % (BA – LA, %)	3.2 (-84.0 ∼ 90.3)	-0.7 (-80.6 ∼ 79.2)	-4.3 (-80.1 ∼ 71.5)

Diff, difference; BA-LA, Bland–Altman 95% limits of agreement.

Discussion

In this study, we demonstrated that IVIM parameters can have good reproducibility when evidential motion-contaminated and/or poorly fitted image data are removed. Several papers have addressed IVIM parameter reproducibility (11 –19,21,23 –29). Those measured and reported reproducibility in comparable formats to our results are listed in Table 6, which also represent the majority of literature on this topic. Our results broadly represent the best results ever reported, though the results for D_fast remain to be further improved in future studies. The difficulty in accurately measuring D_fast has been discussed previously (30). While at this stage we still do not fully know “how good is good enough” regarding the reproducibility, the recent report by Wáng et al. seems to suggest a combined use of IVIM diffusion-weighted imaging (DWI) parameters may have both good sensitivity and good specificity for early stage liver fibrosis detection even when IVIM DWI data acquisition was not optimized (20). Therefore, though the reproducibility for D_fast may not be optimal, there is the possibility that this level of reproducibility may be sufficient for liver fibrosis evaluation. Automated image registration and motion correction for IVIM image series have been reported by other authors (12,18,21,26). However, image registration does not remove artifacts associated with intra-b-value motion and scanner calibration/sequence imperfection. Padhani et al. (5) described the approach for machine and sequence calibrations for DWI. The TE is kept as short as possible as this results in better SNR and it minimizes susceptibility echo planar imaging-related artifacts. Eddy currents related to diffusion gradients and echo planar imaging techniques lead to geometric distortions and image shearing, which can be reduced by increasing readout bandwidths or reducing the echo spacing (5).

Table 6.

Summary of liver IVIM parameter reproducibility in liver parenchyma.

Paper	Subjects	Repeatability or Reproducibility				b-value No.	Breathing and Post processing
Paper	Subjects	Statistics	PF	D_slow	D_fast	b-value No.	Breathing and Post processing
Andreou et al., (2013) (4)	14 Pts (tumor)	BA – LA	-24.3 ∼ 25.1	-5.12 ∼ 8.09	-31.2 ∼ 59.1	8	RT V-by-V LSF
Dyvorne et al., (2013) (12)	6 HV 4 Pts (CLD)	_wCoV (range)	11.5 (1.5 ∼ 22)	4.7 (1.2 ∼ 21.9)	30.1 (9.7 ∼ 50.1)	16	FB ROI Bayesian
Kakite et al., (2015) (15)	11 Pts (CLD)	CoV (range) BA-LA	25.3 (9.3 ∼ 84.9) –74.6 ∼ 56.4	13.2 (1.2 ∼ 28.5) –32.8 ∼ 28.4	59.0 (2.4 ∼ 121.3) –151.2 ∼ 132.7	16	FB ROI Bayesian
Park et al., (2017) (19)	12 HV	_wCoV	9.65	4.05	29.52	9	RT V-by-V LSF (constrains)
Patel et al., (2010) (23)	One Pt (benign lesion) 4 HV	CoV	7.7	3.8	14.6	9	RT ROI Seg-anal/ b=200
Jajamovich et al., (2014) (14)	11 HV 19 Pts (CLD)	CoV BA – LA	20.4 76.9 ∼ 69.0	15.2 –63.5 ∼ 69.7	51.6 –200.5 ∼ 184.5	16	FB ROI Bayesian
Hectors et al., (2016) (26)	8 Pts (CLD)	CoV (range) BA – LA	5.2 (0.0 ∼ 11.8) –20.5 ∼ 21.4	7.3 (0.4 ∼ 19.8) –31.3 ∼ 26.8	21.4 (4.6 ∼ 42.8) –73.8 ∼ 73.7	16	RT and FB ROI Bayesian
Lee et al., (2015) (24)	12 HV	_wCoV (95% CI)	9.86 (7.13 ∼ 12.59)	3.19 (1.83 ∼ 4.54)	71.87 (47.53 ∼ 96.21)	9	ET and RT V-by-V LSF
Gurney-Champion et al., (2016) (21)	intra-session: 14 HV inter-session: 16 HV	_wCoV	Intra-session: 32 Inter-session: 34	Intra-session: 9 Inter-session: 10		14	RT LSF with D_fast fixed

All statistics are in % and the results shown are for liver parenchyma except those of Andreou et al. When different acquisition (FB vs. RT, etc.) or post-processing (full-fitting, segmented fitting, with or without constraint, etc.) schemes were compared, only the best results are presented in this table. Participant number corresponds to the number of participants included for reproducibility analysis.

HV, heathy volunteers, Pts, patients, CLD, chronic liver disease, CoV, mean of the coefficient of variation; _wCoV, within-subject coefficient of variation; BA-LA, Bland–Altman 95% limits of agreement; 95% CI, 95% confidence intervals; RT, free breathing; RT, respiratory triggering (including navigator triggering); ET, electrocardiography triggering; V-by-V, voxel-by-voxel analysis, ROI, ROI analysis; Seg-anal/b = 150, segment fitting method, with b-value threshold of 150 s/mm²; LSF, least square fitting method; Bayesian, Bayesian fitting method.

In addition to the “data cleaning” process taken in this study, a few other steps may have additionally contributed to the good results in this study. The participants were trained to avoid irregular breathing or sudden deep breathing during the examination. Sixteen b-values were used, which is at the upper end of the b-value range compared with published results on IVIM reproducibility (Table 6). We were able to use a median of five slices (average = 5.3 slices) for reproducibility calculation, which is more than the number of slices used in most of the published papers on reproducibility (11 –19,21,23 –29). The signal measurement was a ROI-based method and the IVIM parameters were calculated based on the mean signal intensity of the whole ROI. The ROI-based approach allows for the assessing of the plots of signal measurements and fitted curves for each slices, while this is not possible for pixel-based method when IVIM parameters are generated on parametric maps.

D_slow for liver parenchyma has been shown to have the best reproducibility, followed by PF, and then D_fast. Interestingly, we found a slightly better reproducibility using threshold b-value = 50 or 80 s/mm² for D_slow and PF estimation than threshold b-value = 200 s/mm². On the contrary, a higher b-value threshold (such as 200 s/mm²) can lead to better reproducibility for D_fast (Tables 4 and 5). This may be partially due to more b-values being included for D_slow calculation using a lower b-value threshold. This highlights the importance of acquiring image data with a sufficient b-value number. Wurnig et al. proposed the use of an even lower b-value threshold for liver parenchyma, in the range of 20–40 s/mm², which could reduce the fitting error (31). However, b-value = 200 s/mm² remained the most selected threshold in published literature (7,16,19, 23,32,33), since the signal vs. b-value curve can be assumed to be linear when b-value ≥ 200 s/mm². It should be noted that diffusion compartment is less sensitive to pathologies than perfusion compartment measurement. For example, the study of Luciani et al. (32) reported that D_slow did not differ significantly between healthy and cirrhotic livers, and cirrhotic livers are mainly associated with D_fast reduction. Andreou et al. (4) reported no statistical difference between D_slow of liver parenchyma (=1.1 × 10^–3 mm²/s) and liver metastasis (=1 × 10^–3 mm²/s). Recent literature review also showed that among nine patient studies, only D_fast, despite being the least stable, consistently demonstrated liver fibrosis is progressively associated with a reduced measurement (2). Therefore, it is critically important to measure the perfusion compartment which can be reflected by D_fast, PF, as well as D_slow computed with low threshold b-value to include perfusion element.

Our study has some limitations. The volunteer population in this study included only young healthy participants. While our results may be applicable to diffused liver diseases such as hepatic fibrosis, how our approach can be applicable to focal liver lesions will require additional studies. Second, we did not ask volunteers to fast before examination, while the hepatic flow may vary depending on the fasted/prandial status (34). The reproducibility of IVIM parameters could therefore may be better when the participants are scanned in fasted status. The data cleaning criteria presented in this study remains subjective, not precisely defined, and were not automatized. An objective assessment method, including machine-based recognition of anatomical landmarks and quantitative estimation of scattering of the MRI signal intensity vs. b-value relationship, are being explored in our laboratory to automatically and consistently assess the data acquisition quality. Another limitation is that we only tested segment-unconstrained analysis of IVIM data. While segment-unconstrained analysis remains till now the most popular approach for IVIM analysis (7), it has been suggested that Bayesian probability may perform better in fitting consistency (33,35). Finally, in this study, six out of 54 scanned could not be used for analysis, leading to a success rate of 89%. A better sequence design allowing over-sampling of the focused liver parenchyma regions can minimize the failure rate, and potentially because of the increased number of “sufficient quality slices” available for averaging, will further increase the measure reproducibility. This may be relevant to further improve D_fast scan–rescan reproducibility.

In conclusion, we demonstrated the proof-of-principle that the scan–rescan reproducibility of IVIM parameters can potentially be good. This understanding is important for further developing the IVIM technique for diagnostic clinical application at an individual patient’s level.

Footnotes

Acknowledgments

The authors thank Miss Yao Tina Li, former research student at the Chinese University of Hong Kong, for programming the image processing tool used in this study; Dr Jean-Pierre Cercueil at Department of Vascular and Interventional Radiology, François-Mitterrand Teaching Hospital, University of Bourgogne/Franche-Comté, Dijon, France, for discussions during the course of data analysis; and Dr Weibo Chen, Philips Healthcare Shanghai, China, for setting-up the IVIM diffusion MRI acquisition protocol in Nanjing.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Dr Olivier Chevallier was supported by a grant provided by the Société Française de Radiologie (SFR) together with the Collège des Enseignants de Radiologie de France (CERF).

References

Yamada

Aung

Himeno

et al.

Diffusion coefficients in abdominal organs and hepatic lesions: evaluation with intravoxel incoherent motion echo-planar MR imaging. Radiology 1999; 210: 617–623.

Cercueil

Yuan

et al.

Liver intravoxel incoherent motion (IVIM) magnetic resonance imaging: a comprehensive review of published data on normal values and applications for fibrosis and tumor evaluation. Quant Imaging Med Surg 2017; 7: 59–78.

Koh

Collins

Orton

. Intravoxel incoherent motion in body diffusion-weighted MRI: reality and challenges. Am J Roentgenol 2011; 196: 1351–1361.

Andreou

Koh

Collins

et al.

Measurement reproducibility of perfusion fraction and pseudodiffusion coefficient derived by intravoxel incoherent motion diffusion-weighted MR imaging in normal liver and metastases. Eur Radiol 2013; 23: 428–434.

Padhani

Liu

Mu-Koh

et al.

Diffusion-weighted magnetic resonance imaging as a cancer biomarker: consensus and recommendations. Neoplasia 2009; 11: 102–125.

Turner

Le Bihan

Maier

et al.

Echo-planar imaging of intravoxel incoherent motion. Radiology 1990; 177: 407–414.

Guiu

Petit

J-M

Capitan

et al.

Intravoxel incoherent motion diffusion-weighted imaging in nonalcoholic fatty liver disease: a 3.0-T MR study. Radiology 2012; 265: 96–103.

Huang

Yuan

et al.

Decreases in molecular diffusion, perfusion fraction and perfusion-related diffusion in fibrotic livers: a prospective clinical intravoxel incoherent motion MR imaging study. PLoS One 2014; 9: e113846–e113846.

Yuan

Wong

et al.

Statistical assessment of bi-exponential diffusion weighted imaging signal characteristics induced by intravoxel incoherent motion in malignant breast tumors. Quant Imaging Med Surg 2016; 6: 418–429.

10.

Huang

Zhuo

et al.

Bi-phase age-related brain gray matter magnetic resonance T1ρ relaxation time change in adults. Magn Reson Imaging 2017; 39: 200–205.

11.

Moteki

Horikoshi

. Evaluation of hepatic lesions and hepatic parenchyma using diffusion-weighted echo-planar MR with three values of gradient b-factor. J Magn Reson Imaging 2006; 24: 637–645.

12.

Dyvorne

Galea

Nevers

et al.

Diffusion-weighted imaging of the liver with multiple b values: effect of diffusion gradient polarity and breathing acquisition on image quality and intravoxel incoherent motion parameters-a pilot study. Radiology 2013; 266: 920–929.

13.

Dyvorne

Jajamovich

Kakite

et al.

Intravoxel incoherent motion diffusion imaging of the liver: optimal b-value subsampling and impact on parameter precision and reproducibility. Eur J Radiol 2014; 83: 2109–2113.

14.

Jajamovich

Dyvorne

Donnerhack

et al.

Quantitative liver MRI combining phase contrast imaging, elastography, and DWI: assessment of reproducibility and postprandial effect at 3.0 T. PLoS One 2014; 9: e97355–e97355.

15.

Kakite

Dyvorne

Besa

et al.

Hepatocellular carcinoma: short-term reproducibility of apparent diffusion coefficient and intravoxel incoherent motion parameters at 3.0T. J Magn Reson Imaging 2015; 41: 149–156.

16.

Cohen

Schieke

Hohenwalter

et al.

The effect of low b-values on the intravoxel incoherent motion derived pseudodiffusion parameter in liver. Magn Reson Med 2015; 73: 306–311.

17.

Cui

Dyvorne

Besa

et al.

IVIM diffusion-weighted imaging of the liver at 3.0T: comparison with 1.5T. Eur J Radiol Open 2015; 2: 123–128.

18.

França

Martí-Bonmatí

Alberich-Bayarri

et al.

Evaluation of fibrosis and inflammation in diffuse liver diseases using intravoxel incoherent motion diffusion-weighted MR imaging. Abdom Radiol 2017; 42: 468–477.

19.

Park

Sung

Lee

et al.

Intravoxel incoherent motion diffusion-weighted MRI of the abdomen: The effect of fitting algorithms on the accuracy and reliability of the parameters. J Magn Reson Imaging 2017; 45: 1637–1647.

20.

Wang

YXJ

Deng

et al.

A combined use of intravoxel incoherent motion MRI parameters can differentiate early-stage hepatitis-b fibrotic livers from healthy livers. SLAS Technol 2017. doi: 10.1177/2472630317717049.

21.

Gurney-Champion

Froeling

Klaassen

et al.

Minimizing the acquisition time for intravoxel incoherent motion magnetic resonance imaging acquisitions in the liver and pancreas. Invest Radiol 2016; 51: 211–220.

22.

Cercueil

Petit

Nougaret

et al.

Intravoxel incoherent motion diffusion-weighted imaging in the liver: comparison of mono-, bi- and tri-exponential modelling at 3.0-T. Eur Radiol 2015; 25: 1541–1550.

23.

Patel

Sigmund

Rusinek

et al.

Diagnosis of cirrhosis with intravoxel incoherent motion diffusion MRI and dynamic contrast-enhanced MRI alone and in combination: preliminary experience. J Magn Reson Imaging 2010; 31: 589–600.

24.

Lee

Kim

et al.

Intravoxel incoherent motion diffusion-weighted MR imaging of the liver: effect of triggering methods on regional variability and measurement repeatability of quantitative parameters. Radiology 2015; 274: 405–415.

25.

Shan

Zeng

Liu

et al.

Comparison of free-breathing with navigator-triggered technique in diffusion weighted imaging for evaluation of small hepatocellular carcinoma: effect on image quality and intravoxel incoherent motion parameters. J Comput Assist Tomogr 2015; 39: 709–715.

26.

Hectors

Wagner

Besa

et al.

Intravoxel incoherent motion diffusion-weighted imaging of hepatocellular carcinoma: Is there a correlation with flow and perfusion metrics obtained with dynamic contrast-enhanced MRI?

J Magn Reson Imaging 2016; 44: 856–864.

27.

Barbieri

Donati

Froehlich

et al.

Comparison of intravoxel incoherent motion parameters across MR imagers and field strengths: evaluation in upper abdominal organs. Radiology 2016; 279: 784–794.

28.

Jerome

d’Arcy

Feiweier

et al.

Extended T2-IVIM model for correction of TE dependence of pseudo-diffusion volume fraction in clinical diffusion-weighted magnetic resonance imaging. Phys Med Biol 2016; 61: N667–N680.

29.

Kurugol

Freiman

et al.

Spatially-constrained probability distribution model of incoherent motion (SPIM) for abdominal diffusion-weighted MRI. Med Image Anal 2016; 32: 173–183.

30.

Zhang

Wang

et al.

Cramér-Rao bound for intravoxel incoherent motion diffusion weighted imaging fitting. Conf Proc IEEE Eng Med Biol Soc 2013; 2013: 511–514.

31.

Wurnig

Donati

Ulbrich

et al.

Systematic analysis of the intravoxel incoherent motion threshold separating perfusion and diffusion effects: Proposal of a standardized algorithm. Magn Reson Med. 2015; 74: 1414–1422.

32.

Luciani

Vignaud

Cavet

et al.

Liver cirrhosis: intravoxel incoherent motion MR imaging–pilot study. Radiology 2008; 249: 891–899.

33.

Barbieri

Donati

Froehlich

et al.

Impact of the calculation algorithm on biexponential fitting of diffusion-weighted MRI in upper abdominal organs. Magn Reson Med 2016; 75: 2175–2184.

34.

Pazahr

Nanz

Rossi

et al.

Magnetic resonance imaging of the liver: apparent diffusion coefficients from multiexponential analysis of b values greater than 50 s/mm2 do not respond to caloric intake despite increased portal-venous blood flow. Invest Radiol. 2014; 49: 138–146.

35.

Neil

Bretthorst

. On the use of Bayesian probability theory for analysis of exponential decay data: an example taken from intravoxel incoherent motion experiments. Magn Reson Med 1993; 29: 642–647.

Removal of evidential motion-contaminated and poorly fitted image data improves IVIM diffusion MRI parameter scan–rescan reproducibility

Abstract

Background

Purpose

Material and Methods

Results

Conclusion

Keywords

Introduction

Material and Methods

Results

Discussion

Footnotes

Acknowledgments

Declaration of Conflicting Interests

Funding

References